Thanks for checking the model. Turns out that for me as written above the model would run for me about half the times (giving me the funky results mentioned above), and the other half of the time gave a “bad initial energy” error, I assume for the reasons you give. I am having troubles with models even after exponentiation, but I will continue to tinker with that and report later.
But I wanted to ask you before I go much longer on this, are there are other ways that come to mind to characterize the tightening variance phenomenon? Your answer makes me think that maybe an (exponentiated) linear expression isn’t the only approach here? This negative variance has turned out to be very tricky for me to implement (probably just because I’m new with pymc3). I’ve been sticking to it because I think it well reflects some processes structuring my actual data in a way similar to the graph I showed above.